Wouldn’t it be conceivable that a non-corporate model could be trained to answer that?

Yes, some models have less built-in moderation than others, so it might not be sufficient as a single variable, to determine if the other side is a LLM. But it will trip up some at least.